Microsoft Semantic Kernel CVE-2026-26030 / CVE-2026-25592: prompt injection reaches eval() and the host Startup folder in an AI agent framework, fixed since February
Microsoft shows how two Semantic Kernel flaws turned prompts into shells on the agent host; both were patched in February
Analysis
Microsoft's Defender research team published on 2026-05-07 how two Semantic Kernel flaws turn a prompt injection into code execution on the host running an AI agent (Microsoft Security Blog, 2026-05-07). Both were fixed months earlier: CVE-2026-25592 in February (GitHub Security Advisory GHSA-2ww3-72rp-wpp4, 2026-02-06) and CVE-2026-26030 later the same month (GitHub Security Advisory GHSA-xjw9-4gw8-4rqx, 2026-02-19).
CVE-2026-26030 (Python semantic-kernel before 1.39.4) sits in the In-Memory Vector Store filter behind the Search Plugin: in the default configuration the filter is a Python lambda executed with eval(), and an AI-controlled value is interpolated into it (Microsoft Security Blog, 2026-05-07). Microsoft's exploit prompt escaped the template string, traversed Python's class hierarchy to BuiltinImporter, loaded os and called system(), bypassing the AST-based blocklist; the fix adds node-type and function-call allowlists and blocks class-hierarchy attributes (Microsoft Security Blog, 2026-05-07). The attack needs a prompt-injection vector and the Search Plugin backed by the In-Memory Vector Store in its default configuration (Microsoft Security Blog, 2026-05-07), and the advisory advises against using the In-Memory Vector Store in production (GitHub Security Advisory GHSA-xjw9-4gw8-4rqx, 2026-02-19).
CVE-2026-25592 (.NET Microsoft.SemanticKernel.Plugins.Core before 1.71.0) exposed DownloadFileAsync to the model as a kernel function, so a prompt could have the agent write a script created inside the Azure Container Apps sandbox into the host's Windows Startup folder, where it runs at the next sign-in; upload_file in both SDKs similarly accepted any local path, allowing host files to be read out (Microsoft Security Blog, 2026-05-07). The advisory's workaround is a Function Invocation Filter that allow-lists localFilePath for the download and upload functions (GitHub Security Advisory GHSA-2ww3-72rp-wpp4, 2026-02-06). No source reports exploitation.
Updates1
In CVE-2026-25592, DownloadFileAsync is what writes attacker content to the host, including the Startup folder, and upload_file is what reads host files out (Microsoft Security Blog, 2026-05-07); the earlier text had the two reversed. CVE-2026-26030 passes an interpolated lambda to eval() behind an AST-based blocklist, which Microsoft bypassed through BuiltinImporter (Microsoft Security Blog, 2026-05-07). The fixes were published in February (GitHub Security Advisory GHSA-xjw9-4gw8-4rqx, 2026-02-19) (GitHub Security Advisory GHSA-2ww3-72rp-wpp4, 2026-02-06), three months before Microsoft's write-up.
Sources3
Revision history
- Published 2026-05-10-001
- Correction 2026-10-11T1307Z-audit
The CVE-2026-25592 mechanism was inverted: DownloadFileAsync is the host-side write that drops a script into the Startup folder, and upload_file is the read path. CVE-2026-26030 is an eval() of a lambda with an AST-based blocklist bypassed through BuiltinImporter, not an f-string with a literal-text blocklist. The advisories date from February, not 2026-05-07; the product, EU-pilot and default-store claims, the Python-side blast-radius claim and the EDR category names had no source; and the CVEs need a low-privileged attacker with no user interaction. The entry is rewritten from Microsoft's post and the two advisories with exposure, detection and takeaway lines, a rating and an ATT&CK mapping.
Changed: title headline summary techniques affected_products cves sources classification body deep_dive deep_dive_category
AI-generated · no human review · this permalink is the shareable record for the finding · verify operationally critical claims against the linked primary source.